[PATCH] Don't try to reclose in command_close_bidi_pipe

Some commands require their standard input to be closed (like
git-commit-tree). This patch changes command_close_bidi_pipe so no
longer tries to close already closed handles, resulting in an error.
---
 perl/Git.pm |    1 +
 1 files changed, 1 insertions(+), 0 deletions(-)

diff --git a/perl/Git.pm b/perl/Git.pm
index 7d7f2b1..283bba8 100644
--- a/perl/Git.pm
+++ b/perl/Git.pm

@@ -422,6 +422,7 @@ sub command_close_bidi_pipe {
 	local $?;
 	my ($pid, $in, $out, $ctx) = @_;
 	foreach my $fh ($in, $out) {
+		next unless defined(fileno($fh));
 		unless (close $fh) {
 			if ($!) {
 				carp "error closing pipe: $!";

-- 
1.6.1

Hi,

On Fri, Jan 30, 2009 at 12:59 PM, Yuval Kogman [off-list ref] wrote:

Some commands require their standard input to be closed (like
git-commit-tree). This patch changes command_close_bidi_pipe so no
longer tries to close already closed handles, resulting in an error.

From this message, you seem to intend this as a fix. Can you tell us

how one might go about to reproduce this issue?

Quoting the perldoc for fileno at http://perldoc.perl.org/functions/fileno.html

"(Filehandles connected to memory objects via new features of open may
return undefined even though they are open.)"

Is "unless defined(fileno($fh))" a reliable way to check if the handle
is closed?

-- 
Cheers,
Ray Chuan

Resending, forgot to CC

2009/1/30 Tay Ray Chuan [off-list ref]

From this message, you seem to intend this as a fix. Can you tell us
how one might go about to reproduce this issue?

Here is an example:
http://github.com/yanick/git-cpan-patch/blob/ed67d3f86f371764935fd0da3e7f08536c95b606/git-cpan-import#L190

Since git-commit-tree requires the message to be written before it can
write the sha1 of the new commit object, the handle has to be closed
already, which makes command_close_bidi_pipe die with an error.
The workaround in the linked code reopens a fake handle so that
close() will not error.
For what it's worth, I've been poking at t/t9700/test.pl but since
it's pretty sparse I figured I should contribute this fix
independently.

Is "unless defined(fileno($fh))" a reliable way to check if the handle
is closed?

Yes, since the handle is expected to be the result of a pipe open. The
documentation refers to this case:

    open my $fh, "<", \$buffer;

in which case the filehandle $fh has no associated file descriptor.

fileno() is not a universal check, but it is applicable in this situation.

Hi,

On Fri, Jan 30, 2009 at 11:06 PM, Yuval Kogman [off-list ref] wrote:

Here is an example:
http://github.com/yanick/git-cpan-patch/blob/ed67d3f86f371764935fd0da3e7f08536c95b606/git-cpan-import#L190

Since git-commit-tree requires the message to be written before it can
write the sha1 of the new commit object, the handle has to be closed
already, which makes command_close_bidi_pipe die with an error.
The workaround in the linked code reopens a fake handle so that
close() will not error.
For what it's worth, I've been poking at t/t9700/test.pl but since
it's pretty sparse I figured I should contribute this fix
independently.

Your usage seems to a very special case. You're closing the handle,
and then invoking the method "command_close_bidi_pipe", whose job is
to close the pipe handles passed to it (among other things). So, it
doesn't seem very good to make "command_close_bidi_pipe" to take into
account that you closed your handle independently.

Maybe you could write a pair of new functions for the commit-tree
command in the Git package, something like the
"_open_hash_and_insert_object_if_needed" and "hash_and_insert_object"
pair.

-- 
Cheers,
Ray Chuan